jacobian

commit 6d4caf5d1d9fe9ce3527883e8a5339b2638baa2a
parent 3de195568e32ee4f0d8916d287833259b858a72c
Author: David Freifeld <freifeld.david@gmail.com>
Date:   Sat, 13 Jun 2020 17:12:57 -0700

Backprop is frustrating

Diffstat:
M
test.cpp
 | 
55
+++++++++++++++++++++++++++++++++++++------------------

1 file changed, 37 insertions(+), 18 deletions(-)
diff --git a/test.cpp b/test.cpp
@@ -90,7 +90,7 @@ public:
   int length;
 
   int batch_size;
-  std::vector<int> labels;
+  Eigen::MatrixXd* labels;
 
   Network(char* path, int inputs, int hidden, int outputs, int neurons, int batch_sz);
 
@@ -110,12 +110,13 @@ Network::Network(char* path, int inputs, int hidden, int outputs, int neurons, i
   FILE* fptr = fopen(path, "r");
   int datalen = batch_sz*inputs;
   float batch[datalen];
+  labels = new Eigen::MatrixXd (batch_sz, 1);
   int label;
   char line[1024] = {' '};
   for (int i = 0; i < batch_sz; i++) {
     fgets(line, 1024, fptr);
     sscanf(line, "%f,%f,%f,%f,%i", &batch[0+(i*inputs)], &batch[1+(i*inputs)], &batch[2+(i*inputs)], &batch[3+(i*inputs)], &label);
-    labels.push_back(label);
+    (*labels)(i,0) = label;
   }
   float* batchptr = batch;
   layers.emplace_back(batchptr, batch_sz, inputs);
@@ -132,9 +133,9 @@ Eigen::MatrixXd Network::activate(Eigen::MatrixXd matrix)
 {
   int nodes = matrix.cols();
   for (int i = 0; i < (matrix.rows()*matrix.cols()); i++) {
-    if ((matrix)((float)i / nodes, i%nodes) < 0) {
-      (matrix)((float)i / nodes, i%nodes) = 0;
-    }
+    // if ((matrix)((float)i / nodes, i%nodes) < 0) {
+    //   (matrix)((float)i / nodes, i%nodes) = 0;
+    // }
   }
   return matrix;
 }
@@ -144,7 +145,7 @@ void Network::feedforward()
   for (int i = 0; i < length-1; i++) {
     *layers[i+1].contents = (*layers[i].contents) * (*layers[i].weights);
     for (int j = 0; j < layers[i+1].contents->rows(); j++) {
-      layers[i+1].contents->row(j) += *layers[i+1].bias;
+      // layers[i+1].contents->row(j) += *layers[i+1].bias;
     }
     *layers[i+1].contents = activate(*layers[i+1].contents);
   }
@@ -162,24 +163,41 @@ float Network::cost()
 {
   float sum = 0;
   for (int i = 0; i < layers[length-1].contents->rows(); i++) {
-    sum += pow(labels[i] - (*layers[length-1].contents)(i, 0),2);
+    sum += pow((*labels)(i, 0) - (*layers[length-1].contents)(i, 0),2);
   }
   return (1.0/batch_size) * sum;
 }
 
 float Network::gradient(int mode, int layer, int node)
 {
-  float N = batch_size;
-  if (mode == 0) {
-    for (int i = 0; i < N; i++) {
-      double label = labels[i];
-      double x_i = (layers[layer].contents->row(i) / layers[layer-1].weights->col(i))(0,0) - label;
-      std::
-      // e_i() = e_i - label;
-      // Eigen::MatrixXd w_i = layers[layer-1].weights->col(i);
-      // Eigen::MatrixXd b = layers[layer].bias;
-      // if (w_i.dot(x_i) + b)
-    }
+  // float N = batch_size;
+  // if (mode == 0) {
+  //   for (int i = 0; i < N; i++) {
+  //     double label = labels[i];
+  //     double x_i = (layers[layer].contents->row(i) / layers[layer-1].weights->col(i))(0,0) - label;
+  //     std::
+  //     // e_i() = e_i - label;
+  //     // Eigen::MatrixXd w_i = layers[layer-1].weights->col(i);
+  //     // Eigen::MatrixXd b = layers[layer].bias;
+  //     // if (w_i.dot(x_i) + b)
+  //   }
+  // }
+}
+
+void Network::backpropagate()
+{
+  std::vector<Eigen::MatrixXd> gradients;
+  std::vector<Eigen::MatrixXd> deltas;
+  deltas.push_back(((*layers[length-1].contents - *labels).array() * layers[length-1].contents->array()).matrix());
+  std::cout << deltas[0].matrix() << " \n\n\n " << layers[length-2].contents->transpose() << "\n\n\n\n" << deltas[0].matrix().dot(deltas[0]);
+
+
+
+
+
+  // std::cout << deltas[0];
+  for (int i = length-2; i >= 0; i--) {
+
   }
 }
 
@@ -189,5 +207,6 @@ int main()
   Network net ("./data_banknote_authentication.txt", 4, 2, 1, 5, 10);
   net.feedforward();
   net.list_net();
+  net.backpropagate();
   std::cout << "\nCOST: " << net.cost() << "\n";
 }